[ExI] Vint Cerf on AI, at ORNL

[John Clark]

On Thu, Sep 5, 2019 at 10:06 AM William Flynn Wallace via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> Well, John, knowing the rules is not exactly the case when a situation is
> novel.  Here's what I think should happen:
> The AI plays a program that is highly successful at chess.  The AI is told
> every time it moves in a way that is illegal and of course it notices the
> moves of the other player.  It is told when failure occurs: piece lost,
> checkmate.
> Now that might be really simple for a good AI.  I wouldn't know.  It might
> figure out the rules and scoring in a short period of time and then proceed
> as usual.  But that way at least it is novel.
>

If after 24 hours of self teaching AlphaZero can tell the difference
between a very good Chess move and a superhumanly brilliant Chess move (and
it can) then I don't think it would take it long to tell the difference
between a legal Chess move and a illegal one.

 John K Clark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20190905/fc2ab70a/attachment.htm>